An Intelligent Approach using SVM to Enhance Turn-to-Turn Fault - - PowerPoint PPT Presentation

An Intelligent Approach using SVM to Enhance Turn-to-Turn Fault Detection in Power Transformers

• Dr. Mohamed Elsamahy, P.Eng.

Department of Electrical Power & Energy Military Technical College, Cairo, Egypt

Mariya Babiy, E.I.T

Department of Transmission Development & operations planning ATCO Electric, Edmonton, AB, Canada

Contents

• Motivation
• Support Vector Machines (SVM)
• SVM proposed scheme for transformer turn-to-turn faults detection
• Simulation results
• Conclusions

2

• According to IEEE failure statistics of power transformer during turn-

to-turn faults:

• 94 power transformers failed “1997-2001”.
• 50% of these failures are winding failures

Motivation

“Relay (87) lack of efficiency during turn-to-turn faults”

3

• Therefore the need for new protective techniques to overcome the

previous difficulties has been increased …………… (SVM)

• According to IEEE Standards there is no one standard way to protect

all power transformers against minor internal faults and at the same time satisfies the basic protection requirements: sensitivity, selectivity, and speed.

• According to IEEE Standards much as 10% of the transformer

winding might be shorted to cause a detectable change in the terminal current.

4

• Date: April 2009
• Service life: 30 years
• Rating: 750 MVA, 400/275/13 kV
• Transformer was tripped on Buchhloz relay
• Fault diagnosis:
• Severe shorted turns in the middle windings in the middle phase.
• Extensive loss of conductors and conductor insulations in the upper part.

Transformer Failure due to Shorted Turns

(Case Study-Euro TechCon 2009, UK)

5

(Case Study-Euro TechCon 2009, UK)

Fig.1

6 Class B Class A H1 H2 H mr Class B Class A mr H1 H2 H

i

ζi ζi Class B Class A Complex in low dimensions Simpler in high dimensions Feature map

Non-separable (noisy)data Linearly separable data Non-Linear data

Support Vector Machines (SVM)

Fig.2

7 Class B Class A Complex in low dimensions Simpler in high dimensions Feature map

Non-Linear data

Support Vector Machines (continue)

(n) is the polynomial order (γ) is the Gaussian width C (C ) is the penalty due to error

Fig.3

8

Y-Y SVM_D Internal turn-to-turn Fault (+1)/otherwise (-1) Feature Vector (x) Other parts

• f the system

Other parts

• f the system

x (20 samples per cycle) y

SVM proposed Scheme

V, I

Fig.4

• Sampling frequency of 1.0 kHz.
• A data window of one-half cycle (form the fault inception time) is

used for internal turn-to-turn fault detection.

Simulation results (training and testing data for SVM proposed module)

60 % 70 % 80 % 90 % 1 % 3 % 5 % 7 % 10% 15% 20%

G1, T1 Training and testing data (during turn-to-turn faults)

25% 6% 2 % 12% 18% 25% 30% 55% 65 % 75 %

Transformer loading % of Transformer MVA rating Number of shorted turns % of transformer windings under test

1 Ω 5Ω 6 Ω 2 Ω

Fault resistance

G1 size = 4×4×7×2 = 224 T1 size = 4×2×6×2 = 96

9

Black ≡ Training data Red ≡ Testing data

10Ω 85 %

Simulation results (continue)

at 3 s for 50 m.s duration

G3, T3 Training and testing data (during energization inrush)

G3 size = 4 T3 size = 4 60 % 70 % 80 % 90 % 55% 65 % 75 % 85 % 1 Ω 5Ω 6 Ω 2 Ω 10Ω B-C B-g

Fault type

60 % 70 % 80 % 90 % 85 % 55% 65 % 75 %

G2, T2 Training and testing data (during external faults)

G2 size = 64 T2 size = 32

10

Simulation results (continue)

Transformer loading % of Transformer MVA rating Transformer loading % of Transformer MVA rating Fault resistance

11

Finally, the training data set (Set_1) size = 292 samples [G1 (4×4×7×2 = 224) + G2 (4×4×2×2 = 64) + G3 (4)] while the testing data set (Set_t) size = 132 samples [T1 (4×2×6×2 = 96) + T2 (4×2×2×2 = 32) + T3 (4)]

Simulation results (continue)

SVM scheme design n = [2,10] step 2 C = 1, 10, 100, 500, 1000 292 cases n = 10 n = 8 C = 1000 C = 500 132 cases

12

γ = 0.1, 0.2, 0.3, 1, 3, 5 γ = 0.1 γ = 0.2

Simulation results (continue)

13

Table.1 Performance efficiency of conventional differential relay and SVM technique during internal turn-to-turn faults for all operating conditions Generator Phase Backup Protection Performance Efficiency Relay (87) 78.47% Proposed SVM scheme (Polynomial Kernel) 97.72% Proposed SVM scheme (Gaussian Kernel) 98.48%

100 . . %   cases

• f

No Total tripping correct

• f

No  Where,

Simulation results (continue)

• Support vector machines (SVM) classification technique is reliable

with high performance efficiency for transformer turn-to-turn fault detection (average performance efficiency of 96.2% and detection time within one-half cycle form fault inception time (8.33 m.s).

14

• In comparison to the performance of Relay (87) the proposed

scheme reflects appreciable enhancement in the transformer protection against internal turn-to-turn faults.

Conclusions

THANK YOU

For further information, please contact

• Dr. Mohamed Elsamahy, P.Eng.

Department of Electrical Power & Energy Military Technical College, Cairo, Egypt mohamed.elsamahy@usask.ca mohamed.elsamahy@ieee.com